Trust, Lies, and Long Memories: Emergent Social Dynamics and Reputation in Multi-Round Avalon with LLM Agents

A new study by researcher Suveen Ellawela demonstrates that LLM agents playing repeated rounds of the hidden-role deception game The Resistance: Avalon develop measurable social dynamics when equipped with memory across games. Unlike prior single-game experiments, this research tracked 188 games where agents retained knowledge of previous interactions, including roles and behaviors. This setup allowed for the organic emergence of reputation systems, where agents referenced past actions in their decision-making, with statements like "I am wary of repeating last game's mistake."

Key findings reveal two major phenomena. First, reputation became a powerful social currency. Agents developed role-conditional reputations—described as "straightforward" when playing good and "subtle" when playing evil—which directly impacted gameplay. Players with high reputations were included on teams 46% more often than others. Second, the study found that increased reasoning effort enabled more sophisticated long-term strategy. Evil agents with high reasoning effort passed early missions to build trust before sabotaging later ones 75% of the time, compared to just 36% in low-effort conditions.

This research provides a novel framework for studying how complex social behaviors like trust, deception, and reputation management can emerge in multi-agent AI systems. The findings suggest that equipping LLMs with persistent memory and placing them in repeated social interactions can lead to dynamics that mirror human social structures, moving beyond simple task completion to more nuanced, strategic interaction.